NCVPRIPG 2023 Challenge on Writer Verification (Winning Solution 🌟)

Challenge Webpage: Summer Challenge on Writer Verification, under NCVPRIPG'23
Blog Post: Winning the NCVPRIPG23 Challenge on Writer Verification
Kaggle Link: Summer Challenge on Writer Verification
Wandb logs: wandb
Team Name: MaSha
Summary Paper: assets/NCVPRIPG23_Writer_Verification.pdf

Environment

conda env create -f environment.yml
conda activate ncvp
mkdir -p data/raw

Download and extract the dataset including the val and test CSVs in data/raw.

Download Checkpoints

The checkpoint for the best trained model can be found at Google Drive. You can download this and store it at a desired location. For downloading, you can use gdown.

mkdir pretrained
cd pretrained
gdown 'https://drive.google.com/uc?id=1848Iu-JKXWSBgFvBN50-l-ZXXKgqkRJf'

Generating Results on Test Set

python inference.py --ckpt pretrained/model_best_base.pth.tar

Training

All runs have been logged using wandb

To start training, run

python train.py

You might have to change the inputs to wandb.init() function.

To reproduce the best results, pass the best config

python train.py --config cfgs/best.yml

To disable wandb while training, run

WANDB_MODE=disabled python train.py